Visual long-term memory has a massive storage
capacity for object details
Timothy F. Brady*, Talia Konkle, George A. Alvarez, and Aude Oliva*
Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, Cambridge, MA 02139
object recognition 兩 gist 兩 fidelity
W
e have all had the experience of watching a movie trailer
and having the overwhelming feeling that we can see much
more than we could possibly report later. This subjective experience is consistent with research on human memory, which
suggests that as information passes from sensory memory to
short-term memory and to long-term memory, the amount of
perceptual detail stored decreases. For example, within a few
hundred milliseconds of perceiving an image, sensory memory
confers a truly photographic experience, enabling you to report
any of the image details (1). Seconds later, short-term memory
enables you to report only sparse details from the image (2).
Days later, you might be able to report only the gist of what you
had seen (3).
Whereas long-term memory is generally believed to lack
detail, it is well established that long-term memory can store a
massive number of items. Landmark studies in the 1970s demonstrated that after viewing 10,000 scenes for a few seconds each,
people could determine which of two images had been seen with
83% accuracy (4). This level of performance indicates the
existence of a large storage capacity for images.
However, remembering the gist of an image (e.g., ‘‘I saw a
picture of a wedding not a beach’’) requires the storage of much
less information than remembering the gist and specific details
(e.g., ‘‘I saw that specific wedding picture’’). Thus, to truly
estimate the information capacity of long-term memory, it is
necessary to determine both the quantity of items that can be
remembered and the fidelity (amount of detail), with which each
item is remembered. This point highlights an important limitation of large-scale memory studies (4–6): the level of detail
required to succeed at the memory tests was not systematically
examined. In these studies, the stimuli were images taken from
magazines, where the foil items used in the two-alternative
forced-choice tests were random images drawn from the same set
(4). Thus, the foil items were typically quite different from the
www.pnas.org兾cgi兾doi兾10.1073兾pnas.0803390105
studied images, making it impossible to conclude whether the
memories for each item in these previous experiments consisted
of only the ‘‘gist’’ or category of the image, or whether they
contained specific details about the images. Therefore, it remains unclear exactly how much visual information can be stored
in human long-term memory.
There are reasons for thinking that the memories for each item
in these large-scale experiments might have consisted of only the
gist or category of the image. For example, a well known body
of research has shown that human observers often fail to notice
significant changes in visual scenes; for instance, if their conversation partner is switched to another person, or if large
background objects suddenly disappear (7, 8). These ‘‘change
blindness’’ studies suggest that the amount of information we
remember about each item may be quite low (8). In addition, it
has been elegantly demonstrated that the details of visual
memories can easily be interfered with by experimenter suggestion, a matter of serious concern for eyewitness testimony, as well
as another indication that visual memories might be very sparse
(9). Taken together, these results have led many to infer that the
representations used to remember the thousands of images from
the experiments of Shepard (5) and Standing (4) were in fact
quite sparse, with few or no details about the images except for
their basic-level categories (8, 10–12).
However, recent work has also suggested that visual long-term
memory representations can be more detailed than previously
believed. Long-term memory for objects in scenes can contain
more information than only the gist of the object (13–16). For
instance, Hollingworth (13) showed that, when requiring memory for a hundred or more objects, observers remain significantly
above chance at remembering which exemplar of an object they
have seen (e.g., ‘‘did you see this power drill or that one?’’) even
after seeing up to 400 objects in between studying the object and
being tested on it. This result suggests that memory is capable of
storing fairly detailed visual representations of objects over long
time periods (e.g., longer than working memory).
The current study was designed to estimate the information
capacity of visual long-term memory by simultaneously pushing
the system in terms of both the quantity and the fidelity of the
representations that must be stored. First, we used isolated
objects that were not embedded in scenes, to more systematically
control the conceptual content of the stimulus set and prevent
the role of contextual cues that may have contributed to memory
performance in previous experiments. In addition, we used very
subtle visual discriminations to probe the fidelity of the visual
representations. Last, we had people remember several thouAuthor contributions: T.F.B., T.K., G.A.A., and A.O. designed research, performed research,
analyzed data, and wrote the article.
The authors declare no conflict of interest.
This article is a PNAS Direct Submission.
Freely available online through the PNAS open access option.
*To whom correspondence may be addressed. E-mail: tfbrady@mit.edu or oliva@mit.edu.
This article contains supporting information online at www.pnas.org/cgi/content/full/
0803390105/DCSupplemental.
© 2008 by The National Academy of Sciences of the USA
PNAS 兩 September 23, 2008 兩 vol. 105 兩 no. 38 兩 14325–14329
PSYCHOLOGY
One of the major lessons of memory research has been that human
memory is fallible, imprecise, and subject to interference. Thus,
although observers can remember thousands of images, it is
widely assumed that these memories lack detail. Contrary to this
assumption, here we show that long-term memory is capable of
storing a massive number of objects with details from the image.
Participants viewed pictures of 2,500 objects over the course of
5.5 h. Afterward, they were shown pairs of images and indicated
which of the two they had seen. The previously viewed item could
be paired with either an object from a novel category, an object of
the same basic-level category, or the same object in a different
state or pose. Performance in each of these conditions was remarkably high (92%, 88%, and 87%, respectively), suggesting that
participants successfully maintained detailed representations of
thousands of images. These results have implications for cognitive
models, in which capacity limitations impose a primary computational constraint (e.g., models of object recognition), and pose a
challenge to neural models of memory storage and retrieval, which
must be able to account for such a large and detailed storage
capacity.
NEUROSCIENCE
Edited by Dale Purves, Duke University Medical Center, Durham, NC, and approved August 1, 2008 (received for review April 8, 2008)
Fig. 1. Example test pairs presented during the two-alternative forced-choice task for all three conditions (novel, exemplar, and state). The number of observers
reporting the correct item is shown for each of the depicted pairs. The experimental stimuli are available from the authors.
sand objects. Combined, these manipulations enable us to
estimate a new bound on the capacity of memory to store visual
information.
Results
Observers were presented with pictures of 2,500 real world
objects for 3 s each. The experiment instructions and displays
were designed to optimize the encoding of object information
into memory. First, observers were informed that they should try
to remember all of the details of the items (17). Second, objects
from mostly distinct basic-level categories were chosen to minimize conceptual interference (18). Last, memory was tested
with a two-alternative forced-choice test, in which a studied item
was paired with a foil and the task was to choose the studied item,
allowing for recognition memory rather than recall memory (as
in 4).
We varied the similarity of the studied item and the foil item
in three ways (Fig. 1). In the novel condition, the old item was
paired with a new item that was categorically distinct from all of
the previously studied objects. In this case, remembering the
category of the object, even without remembering the visual
14326 兩 www.pnas.org兾cgi兾doi兾10.1073兾pnas.0803390105
details of the object, would be sufficient to choose the appropriate item. In the exemplar condition, the old item was paired
with a physically similar new object from the same basic-level
category. In this condition, remembering only the basic-level
category of the object would result in chance performance. Last,
in the state condition, the old item was paired with a new item
that was exactly the same object, but appeared in a different state
or pose. In this condition, memory for the category of the object,
or even for the identity of the object, would be insufficient to
select the old item from the pair. Thus, memory for specific
details from the image is required to select the appropriate
object in both the exemplar and state conditions. Critically,
observers did not know during the study session which items of
the 2,500 would be tested afterward, nor what they would be
tested against. Thus, any strategic encoding of a specific detail
that would distinguish between the item and the foil was not
possible. To perform well on average in both the exemplar and
the state conditions, observers would have to encode many
specific details from each object.
Performance was remarkably high in all three of the test
conditions in the two-alternative forced choice (Fig. 2). As
Brady et al.
anticipated based on previous research (4), performance was
high in the novel condition, with participants correctly reporting
the old item on 92.5% (SEM, 1.6%) of the trials. Surprisingly,
performance was also exceptionally high in both conditions that
required memory for details from the images: on average,
participants responded correctly on 87.6% (SEM, 1.8%) of
exemplar trials, and 87.2% (SEM, 1.8%) of the state trials. A
one-way repeated measures ANOVA revealed a significant
effect of condition, F(2,26) ⫽ 11.3, P ⬍ 0.001. Planned pairwise
t tests show that performance in the novel condition was
significantly more accurate than the state and exemplar conditions [novel vs. exemplar: t(13) ⫽ 3.4, P ⬍ 0.01; novel vs. state:
t(13) ⫽ 4.3, P ⬍ 0.01; and exemplar vs. state, n.s. P ⬎ 0.10].
However, reaction time data were slowest in the state condition,
intermediate in the exemplar condition, and fastest in the novel
conditions [M ⫽ 2.58 s, 2.42 s, 2.28 s, respectively; novel vs.
exemplar: t(13) ⫽ 1.81, P ⫽ 0.09; novel vs. state: t(13) ⫽ 4.05, P ⫽
0.001; and exemplar vs. state: t(13) ⫽ 2.71, P ⫽ 0.02], consistent
with the idea that the novel, exemplar, and state conditions
required increasing detail. Participant reports afterward indicated that they were usually explicitly aware of which item they
had seen, as they expressed confidence in their performance and
volunteered information about the details that enabled them to
pick the correct items.
During the presentation of the 2,500 objects, participants
monitored for any repeat images. Unbeknownst to the participants, these repeats occurred anywhere from 1 to 1,024 images
previously in the sequence (on average, one in eight images was
a repeat). This task insured that participants were actively
attending to the stream of images as they were presented and
provided an online estimate of memory storage capacity over the
course of the entire study session.
Performance on the repeat-detection task also demonstrated
remarkable memory. Participants rarely false-alarmed (1.3%;
SEM, ⫾ 1%), and were highly accurate in reporting actual
repeats (96% overall; SEM, ⫾ 1%). Accuracy was near ceiling
for repeat images with up to 63 intervening items, and declined
gradually for detecting repeat items with more intervening items
(supporting information (SI) Text and Fig. S1). Even at the
longest condition of 1,023 intervening items (i.e., items that were
initially presented ⬇2 h previously), the repeats were detected
⬇80% of the time. The correlation between the performance of
the observers at the repeat-detection task and their performance
in the forced choice was high (r ⫽ 0.81). The error rate as a
function of number of intervening items fits well with a standard
Brady et al.
The Information Capacity of Memory. Memory capacity cannot
solely be characterized by the number of items stored: a proper
capacity estimate takes into account the number of items
remembered and multiplies this by the amount of information
per item. In the present experiment we show that the information
remembered per item is much higher than previously believed,
as observers can correctly choose among visually similar foils.
Therefore, any estimate of long-term memory capacity will be
significantly increased by the present data. Ideally, we could
quantify this increase, for example, by using informationtheoretic bits, in terms of the actual visual code used to represent
the objects. Unfortunately, one must know how the brain
encodes visual information into memory to truly quantify capacity in this way.
However, Landauer (20) provided an alternate method for
quantifying the capacity of memory by calculating the number of
bits required to correctly make a decision about which items have
been seen and which have not (21). Rather than assign images
codes based on visual similarity, this model assigns each image
a random code regardless of its visual appearance. Memory
errors happen when two images are assigned the same code. In
this model the optimal code length is computed from the total
number of items to remember and the percentage correct
achieved on a two-alternative forced-choice task (see SI Text).
Importantly, this model does not take into account the content
of the remembered items: the same code length would be
obtained if people remembered 80% of 100 natural scenes or
80% of 100 colored letters. In other words, the bits in the model
refer to content-independent memory addresses, and not estimated codes used by the visual system.
Given the 93% performance in the novel condition, the
optimal code would require 13.8 bits per item, which is comparable with estimates of 10–14 bits needed for previous large-scale
experiments (20). To expand on the model of Landauer, we
assume a hierarchical model of memory where we first specify
the category and the additional bits of information specify the
exemplar and state of the item in that category (see SI Text). To
match 88% performance in the exemplar conditions, 2.0 additional bits per item are required for each item. Similarly, 2.0
PNAS 兩 September 23, 2008 兩 vol. 105 兩 no. 38 兩 14327
PSYCHOLOGY
Fig. 2. Memory performance for each of the three test conditions (novel,
exemplar, and state) is shown above. Error bars represent SEM. The dashed
line indicates chance performance.
Discussion
We found that observers could successfully remember details
about thousands of images after only a single viewing. What do
these data say about the information capacity of visual long-term
memory? It is known from previous research that people can
remember large numbers of pictures (4–6) but it has often been
assumed that they were storing only the gist of these images (8,
10–12). Whereas some evidence suggested observers are capable
of remembering details about a few hundred objects over long
time periods (13), to our knowledge, no experiment had previously demonstrated accurate memory at the exemplar or state
level on such a large scale. The present results demonstrate visual
memory is a massive store that is not exhausted by a set of 2,500
detailed representations of objects. Importantly, these data
cannot reveal the format of these representations, and should
not be taken to suggest that observers have a photographic-like
memory (19). Further work is required to understand how the
details from the images are encoded and stored in visual
long-term memory.
NEUROSCIENCE
power law of forgetting (r2 ⫽ 0.98). The repeat-detection task
also shows that this high capacity memory arises not only in
two-alternative forced-choice tasks, but also in ongoing old/new
recognition tasks, although the repeat-detection task did not
probe for more detailed representations beyond the category
level. Together with the memory test, these results indicate a
massive capacity-memory system, in terms of both the quantity
and fidelity of the visual information that can be remembered.
additional bits are required to achieve 87% correct in the state
condition. Thus, we increase the estimated code length from 13.8
to 17.8 bits per item. This raises the lower bound on our estimate
of the representational capacity of long-term memory by an
order of magnitude, from ⬇14,000 (213.8) to ⬇228,000 (217.8)
unique codes. This number does not tell us the true visual
information capacity of the system. However, this model is a
formal way of demonstrating how quickly any estimate of
memory capacity grows if we increase the size of the representation of each individual object in memory.
Why examine the capacity of people to remember visual
information? One reason is that the evolution of more complicated cognition and behavioral repertoires involved the gradual
enlargement of the long-term memory capacities of the brain
(22). In particular, there are reasons to believe the capacity of
our memory systems to store perceptual information may be a
critical factor in abstract reasoning (23, 24). It has been argued,
for example, that abstract conceptual knowledge that appears
amodal and abstracted from actual experience (25) may in fact
be grounded in perceptual knowledge (e.g., perceptual symbol
systems; see ref. 26). Under this view, abstract conceptual
properties are created on the fly by mental simulations on
perceptual knowledge. This view suggests an adaptive significance for the ability to encode a large amount of information in
memory: storing large amounts of perceptual information allows
abstraction based on all available information, rather than
requiring a decision about what information might be necessary
at some later point in time (24, 27).
Organization of Memory. All 2,500 items in our study stream were
categorically distinct and thus had different high level, conceptual representations. Long-term memory is often seen as organized by conceptual similarity (e.g., in spreading activation
models; see refs. 28 and 29). Thus, the conceptual distinctiveness
of the objects may have reduced interference between them and
helped support the remarkable memory performance we observed (18). In addition, recent work has suggested that the
representation of the perceptual features of an object may often
differ depending on the category the object is drawn from (30).
Taken together, these ideas suggest an important role for
categories and concepts in the storage of the visual details of
objects, an important area of future research.
Another possible distinction in the organization of memory is
between memory for objects, memory for collections of objects,
and memory for scenes. Whereas some work has shown that it
is possible to remember details from scenes drawn from the same
category (15), future work is required to examine massive and
detailed memory for complex scenes.
Familiarity versus Recollection. The literature on long-term mem-
ory frequently distinguishes between two types of recognition
memory: familiarity, the sense that you have seen something
before; and recollection, specific knowledge of where you have
seen it (31). However, there remains controversy over the extent
to which these types of memory can be dissociated, and the
extent to which forced-choice judgments tap into familiarity
more than recollection or vice versa (32). In addition, whereas
some have argued that perceptual information is often more
associated with familiarity and conceptual information is associated more with recollection (31), this view also remains
disputed (32, 33). Thus, it is unclear the relative extent to which
choices of the observers in the current two-alternative forcedchoice tests were based on familiarity versus recollection. Given
the perceptual nature of the details required to select the studied
item, it is likely that familiarity has a major role, and that
recollection aids recognition performance on the subset of trials
in which observers were explicitly aware of the details that were
most helpful to their decision (a potentially large subset of trials,
14328 兩 www.pnas.org兾cgi兾doi兾10.1073兾pnas.0803390105
based on self-reports). Importantly, however, whether observers
were depending on recollection or familiarity alone, the stored
representation still requires enough detail to distinguish it from
the foil at test. Our main conclusion is the same whether the
memory is subserved by familiarity or recollection: observers
encoded and retained many specific details about each object.
Constraints on Models of Object Recognition and Categorization.
Long-term memory capacity imposes a constraint on high-level
cognitive functions and on neural models of such functions. For
example, approaches to object recognition often vary by either
relying on brute force online processing or a massive parallel
memory (34, 35). The present data lend credence to objectrecognition approaches that require massive storage of multiple
object viewpoints and exemplars (36–39). Similarly, in the
domain of categorization, a popular class of models, so-called
exemplar models (40), have suggested that human categorization
can best be modeled by positing storage of each exemplar that
is viewed in a category. The present results demonstrate the
feasibility of models requiring such large memory capacities.
In the domain of neural models, the present results imply that
visual processing stages in the brain do not, by necessity, discard
visual details. Current models of visual perception posit a
hierarchy of processing stages that reach more and more abstract
representations in higher-level cortical areas (35, 41). Thus, to
maintain featural details, long-term memory representations of
objects might be stored throughout the entire hierarchy of the
visual processing stream, including early visual areas, possibly
retrieved on demand by means of a feedback process (41, 42).
Indeed, imagery processes, a form of representation retrieval,
have been shown to activate both high-level visual cortical areas
and primary visual cortex (43). In addition, functional MRI
studies have indicated that a relatively mid-level visual area, the
right fusiform gyrus, responds more when observers are encoding objects for which they will later remember the specific
exemplar, compared with objects for which they will later
remember only the gist (44). Understanding the neural substrates underlying this massive and detailed storage of visual
information is an important goal for future research and will
inform the study of visual object recognition and categorization.
Conclusion
The information capacity of human memory has an important
role in cognitive and neural models of memory, recognition, and
categorization, because models of these processes implicitly or
explicitly make claims about the level of detail stored in memory.
Detailed representations allow more computational flexibility
because they enable processing at task-relevant levels of abstraction (24, 27), but these computational advantages tradeoff with
the costs of additional storage. Therefore, establishing the
bounds on the information capacity of human memory is critical
to understanding the computational constraints on visual and
cognitive tasks.
The upper bound on the size of visual long-term memory has
not been reached, even with previous attempts to push the
quantity of items (4), or the attempt of the present study to push
both the quantity and fidelity. Here, we raise only the lower
bound of what is possible, by showing that visual long-term
memory representations can contain not only gist information
but also details sufficient to discriminate between exemplars and
states. We think that examining the fidelity of memory representations is an important addition to existing frameworks of
visual long-term memory capacity. Whereas in everyday life we
may often fail to encode the details of objects or scenes (7, 8, 17),
our results suggest that under conditions where we attempt to
encode such details, we are capable of succeeding.
Brady et al.
Stimuli. Stimuli were gathered by using both a commercially available database
(Hemera Photo-Objects, Vol. I and II) and internet searches by using Google Image
Search. Overall, 2,600 categorically distinct images were gathered for the main
database, plus 200 paired exemplar images and 200 paired state images drawn
from categories not represented in the main database. The experimental stimuli
are available from the authors. Once these images had been gathered, 200 were
selected at random from the 2,600 objects to serve in the novel test condition.
Thus, all participants were tested with the same 300 pairs of novel, exemplar, and
state images. However, the item seen during the study session and the item used
as the foil at test were randomized across participants.
Study Blocks. The experiment was broken up into 10 study blocks of ⬇20 min
each, followed by a 30 min of testing session. Between blocks participants were
given a 5-min break, and were not allowed to discuss any of the images they had
seen. During a block, ⬇300 images were shown, with 2,896 images shown overall:
2,500 new and 396 repeated images. Each image (subtending 7.5 by 7.5° of visual
angle) was presented for 3 s, followed by an 800-ms fixation cross.
Repeat-Detection Task. To maintain attention and to probe online memory
capacity, participants performed a repeat-detection task during the 10 study
blocks. Repeated images were inserted into the stream such that there were
between 0 and 1,023 intervening items, and participants were told to respond by
using the spacebar anytime that an image repeated throughout the entire study
period. They were not informed of the structure of the repeat conditions.
Participants were given feedback only when they responded, with the fixation
cross turning red if they had incorrectly pressed the space bar (false alarm) or
green if they had correctly detected a repeat (hit), and were given no feedback
for misses or correct rejections.
1. Sperling G (1960) The information available in brief visual presentations. Psychol
Monogr 74:1–29.
2. Phillips WA (1974) On the distinction between sensory storage and short-term visual
memory. Percept Psychophys 16:283–290.
3. Brainerd CJ, Reyna VF (2005) The Science of False Memory (Oxford Univ Press, New
York)
4. Standing L (1973) Learning 10,000 pictures. Q J Exp Psychol 25:207–222.
5. Shepard RN (1967) Recognition memory for words, sentences, and pictures. J Verb
Learn Verb Behav 6:156 –163.
6. Standing L, Conezio J, Haber RN (1970) Perception and memory for pictures: Single-trial
learning of 2500 visual stimuli. Psychon Sci 19:73–74.
7. Rensink RA, O’Regan JK, Clark JJ (1997) To see or not to see: The need for attention to
perceive changes in scenes. Psychol Sci 8:368 –373.
8. Simons DJ, Levin DT (1997) Change blindness. Trends Cogn Sci 1:261–267.
9. Loftus EF (2003) Our changeable memories: Legal and practical implications. Nat Rev
Neurosci 4:231–234.
10. Chun MM (2003) in Cognitive Vision. Psychology of Learning and Motivation: Advances
in Research and Theory, eds Irwin D, Ross BH (Academic, San Diego, CA), Vol 42, pp
79 –108.
11. Wolfe JM (1998) Visual Memory: What do you know about what you saw? Curr Biol
8:R303–R304.
12. O’Regan JK, Noë A (2001) A sensorimotor account of vision and visual consciousness.
Behav Brain Sci 24:939 –1011.
13. Hollingworth A (2004) Constructing visual representations of natural scenes: The roles
of short- and long-term visual memory. J Exp Psychol Hum Percept Perform 30:519 –
537.
14. Tatler BW, Melcher D (2007) Pictures in mind: Initial encoding of object properties
varies with the realism of the scene stimulus. Perception 36:1715–1729.
15. Vogt S, Magnussen S (2007) Long-term memory for 400 pictures on a common theme.
Exp Psycol 54:298 –303.
16. Castelhano M, Henderson J (2005) Incidental visual memory for objects in scenes. Vis
Cog 12:1017–1040.
17. Marmie WR, Healy AF (2004) Memory for common objects: Brief intentional study is
sufficient to overcome poor recall of US coin features. Appl Cognit Psychol 18:445– 453.
18. Koutstaal W, Schacter DL (1997) Gist-based false recognition of pictures in older and
younger adults. J Mem Lang 37:555–583.
19. Intraub H, Hoffman JE (1992) Remembering scenes that were never seen: Reading and
visual memory. Am J Psychol 105:101–114.
20. Landauer TK (1986) How much do people remember? Some estimates of the quantity
of learned information in long-term memory. Cognit Sci 10:477– 493.
21. Dudai Y (1997) How big is human memory, or on being just useful enough. Learn Mem
3:341–365.
Brady et al.
Overall, 56 images were repeated immediately (1-back), 52 were repeated
with 1 intervening item (2-back), 48 were repeated with 3 intervening items
(4-back), 44 were repeated with 7 intervening items (8-back), and so forth, down
to 16 repeated with 1,023 intervening items (1,024-back). Repeat items were
inserted into the stream uniformly, with the constraint that all of the lengths of
n-backs (1-back, 2-back, 4-back, and 1,024-back) had to occur equally in the first
half of the experiment and the second half. This design ensured that fatigue
would not differentially affect images that were repeated from further back in
the stream. Due to the complexity of generating a properly counterbalanced set
of repeats, all participants had repeated images appear at the same places within
the stream. However, each participant saw a different order of the 2,500 objects,
and the specific images repeated in the n-back conditions were also different
across participants. Images that would later be tested in one of the three memory
conditions were never repeated during the study period.
Forced-Choice Tests. Following a 10-min break after the study period, we
probed the fidelity with which objects were remembered. Two items were
presented on the screen, one previously seen old item, and one new foil item.
Observers reported which item they had seen before in a two-alternative
forced-choice task.
Participants were allowed to proceed at their own pace and were told to
emphasize accuracy, not speed, in making their judgments. The 300 test trials
were presented in a random order for each participant, with the three types
of test trials (novel, exemplar, and state) interleaved. The images that would
later be tested were distributed uniformly throughout the study period.
ACKNOWLEDGMENTS. We thank P. Cavanagh, M. Chun, M. Greene, A.
Hollingworth, G. Kreiman, K. Nakayama, T. Poggio, M. Potter, R. Rensink, A.
Schachner, T. Thompson, A. Torralba, and J. Wolfe for helpful conversation
and comments on the manuscript. This work was partly funded by National
Institutes of Health Training Grant T32-MH020007 (to T.F.B.), a National
Defense Science and Engineering Graduate Fellowship (T.K.), National Research Service Award Fellowship F32-EY016982 (to G.A.A.), and National
Science Foundation (NSF) Career Award IIS-0546262 and NSF Grant IIS0705677 (to A.O.).
22. Fagot J, Cook RG (2006) Evidence for large long-term memory capacities in baboons
and pigeons and its implications for learning and the evolution of cognition. Proc Natl
Acad Sci USA 103:17564 –17567.
23. Fuster JM (2003) More than working memory rides on long-term memory. Behav Brain
Sci 26:737.
24. Palmeri TJ, Tarr MJ (2008) Visual Memory, eds Luck SJ, Hollingworth A (Oxford Univ
Press, New York), 163–207.
25. Fodor JA (1975). The Language of Thought (Harvard Univ Press, Cambridge, MA).
26. Barsalou LW (1999) Perceptual symbol systems. Behav Brain Sci 22:577– 660.
27. Barsalou LW (1990) Content and Process Specificity in the Effects of Prior Experiences,
eds Srull TK, Wyer RS, Jr (Erlbaum, Hillsdale, NJ), pp 61– 88.
28. Collins AM, Loftus EF (1975) A spreading activation theory of semantic processing.
Psychol Rev 82:407– 428.
29. Anderson JR (1983) Retrieval of information from long-term memory. Science 220:25–30.
30. Sigala N, Gabbiani F, Logothetis NK (2002) Visual categorization and object representation in monkeys and humans. J Cog Neuro 14:187–198.
31. Mandler G (1980) Recognizing: The judgment of previous occurrence. Psychol Rev
87:252–271.
32. Yonelinas AP (2002) The nature of recollection and familiarity: A review of 30 years of
research. J Mem Lang 46:441–517.
33. Jacoby LL (1991) A process dissociation framework: Separating automatic from intentional uses of memory. J Mem Lang 30:513–541.
34. Tarr MJ, Bulthoff HH (1998) Image-based object recognition in man, monkey and
machine. Cognition 67:1–20.
35. Riesenhuber M, Poggio T (1999) Hierarchical models of object recognition in cortex.
Nat Neurosci 2:1019 –1025.
36. Logothetis NK, Pauls J, Poggio T (1995) Shape representation in the inferior temporal
cortex of monkeys. Curr Biol 5:552–563.
37. Tarr MJ, Williams P, Hayward WG, Gauthier, I. (1998) Three-dimensional object recognition is viewpoint dependent. Nat Neurosci 1:275–277.
38. Bulthoff HH, Edelman S (1992) Psychological support for a two-dimensional view
interpolation theory of object recognition. Proc Natl Acad Sci USA 89:60 – 64.
39. Torralba A, Fergus R, Freeman W, 80 million tiny images: A large dataset for nonparametric object and scene recognition. IEEE Trans PAMI, in press.
40. Nosofsky RM (1986) Attention, similarity, and the identification-categorization relationship. J Exp Psychol Gen 115:39 – 61.
41. Ahissar M, Hochstein S (2004) The reverse hierarchy theory of visual perceptual
learning. Trends Cogn Sci 8:457– 464.
42. Wheeler ME, Petersen SE, Buckner RL (2000) Memory’s echo: Vivid remembering
reactivates sensory-specific cortex. Proc Natl Acad Sci USA 97:11125–11129.
43. Kosslyn SM, et al. (1999) The Role of Area 17 in Visual Imagery: Convergent Evidence
from PET and rTMS. Science 284:167–170.
44. Garoff RJ, Slotnick SD, Schacter DL (2005) The neural origins of specific and general
memory: The role of fusiform cortex. Neuropsychologia 43:847– 859.
PNAS 兩 September 23, 2008 兩 vol. 105 兩 no. 38 兩 14329
NEUROSCIENCE
Participants. Fourteen adults (aged 20 –35) gave informed consent and participated in the experiment. All of the participants were tested simultaneously, by using computer workstations that were closely matched for
monitor size and viewing distance.
PSYCHOLOGY
Materials and Methods
Fig. S1. Performance on detecting repeat images during the 5.5 h study session. Images were repeated with a different number of intervening items, from
0 to 1,023, by powers of 2.
Brady et al. www.pnas.org/cgi/content/short/0803390105
2 of 2
Supporting Information
Brady et al. 10.1073/pnas.0803390105
SI Text
Modeling Methods. Our model of memory capacity in bits per item
is based on work by Landauer (1), who provided a method for
estimating the number of bits required to correctly make a
decision about which items have been seen and which have not.
This simple model assigns each picture a random code (b bits
long), rather than assign them based on visual similarity. If a foil
item is assigned the same code as any old item, the old item
cannot be distinguished from the foil. Because this is a twoalternative forced-choice task, an error will occur on one half of
these cases. Given this model we can compute the length of code
(in bits) necessary to achieve any given accuracy level for a given
number of items.
For example, to achieve 88% correct with 1,000 items in
memory, you must have a code at least 11.9 bits long (3,821
possible codes). The chance of a new item being assigned a code
that overlaps an old item would then be 23% ([1 ⫺ 1/3,821]1,000),
and on half of trials observers would still answer correctly,
resulting in 12% errors, or 88% correct performance. In general,
the number of bits is related to memory performance in this
model by the following equation: b ⫽ ⫺log2[1 ⫺ (2p ⫺ 1)1/n],
where p is the percentage correct, and n is the number of items
in memory.
If this model is capturing a systematic property of memory,
similar estimates for the number of bits should be obtained by
using different numbers of pictures in memory (because performance should increase correspondingly). Supporting this, Landauer (1) found that similar calculations result in estimates of 10.0
bits with 20 pictures in memory and 99% correct judgments, 10.2
bits with 400 pictures in memory and 86% correct judgments, etc.
By using this model, we find observers must have 13.8 bits of
information per item based on our novel condition (92% correct
with 2,500 pictures in memory). This suggests that the maximum
capacity of memory, if all items were coded by using the optimal
set of features (decision-level bits), would be 213.8 (14,000)
unique items.
To model the exemplar and state results, we make the
assumption that memory is organized hierarchically, such that
the bits for the category appear before the bits for the exemplar
per state, resulting in greater similarity in the codes (and greater
chance of confusion) for items within a category than items in
different categories. In the exemplar condition, observers are
holding one item of the category in memory and this results in
⬇87.5% correct. If observers have two bits of exemplar-level
information about each studied item, then the probability of the
a foil exemplar overlapping with a studied item would be 25%
(1/22), resulting in 12.5% error trials, thus 87.5% correct performance, matching our empirical results. The same logic holds
in the state condition which also has ⬇87.5% performance.
Thus, our overall estimate of memory capacity is 17.8 bits per
item, where 13.8 bits are required to code the category of the
object, and the additional four bits per item are required to code
which exemplar (two bits) and state (two bits) the object is to
successfully distinguish it from the foils. This suggests that
maximum number of unique items that can be put in memory
(assuming an optimal feature set) is 217.8, or 228,000.
This model, in which the exemplar bits are separate from the
category bits, is more conservative than giving unique codes
without regard to category, since it accounts for the idea that we
are more likely to confuse two teacups than a teacup and a
tractor. However, our calculations do assume that the exemplar
and state conditions draw on different bits, such that the
information used to perform well in the exemplar tests is not the
same as the information used for the state tests. This is compatible with memory representations in which it is possible to
know that you saw an open door rather than a closed door (state
condition) without knowing exactly which door it was (exemplar
condition). However, even if up to half of the information was
shared between the exemplar and state condition, we would still
obtain an estimate of 16.8 bits of information, or 114,000 unique
codes, still an order of magnitude over previous estimates.
A previous study by Hollingworth (2) also examined object
representations on the order of hundreds of images. Participants
were shown scenes with many embedded objects and were
subsequently tested with exemplar-level foil items. To quantify
the capacity of memory estimated in this experiment, we used the
same hierarchical model. This study did not include a novel test
condition to estimate the category bits, so we assumed a generous 99% performance, giving 14.27 bits. Performance in the
exemplar level tests was 65%, or an additional 0.5 bits. Thus we
estimate the memory capacity demonstrated in Hollingworth (2)
between 14–15 bits, approximately equal to the estimates arrived
at by Landauer (1). Interestingly, this study demonstrates that
memory can store hundreds of objects with exemplar-level
fidelity, and even this does not guarantee an increased estimate
of memory capacity.
The hierarchical decision-level model is based on the number
of items studied, and importantly, on the number of questions
asked about each item. Previous large-scale memory studies only
tested against a novel foil (one question). Here, we ask about
novel, exemplar, and state comparisons (three questions, thus
three sets of bits). Our estimate of memory capacity could be
increased even further if more questions were asked about what
kind of information observers have about remembered items.
However, there is an important caution to this modeling approach: the questions asked about each item are probably not
independent. For example, we could have included a fourth kind
of test asking about the orientation of the presented object, and
added more bits to the estimated code of each item. However,
information about the state is probably also informative about
orientation. Thus, asking 1,000 different questions to show an
enormous memory capacity estimate is not sufficient because
such questions will likely have overlapping information when
considered under the true coding model of the visual system. On
a positive note, if one could ask the right set of completely
independent questions, one might be approximating the visual
coding scheme.
1. Landauer TK (1986) How much do people remember? Some estimates of the quantity
of learned information in long-term memory. Cognit Sci 10:477– 493.
2. Hollingworth A (2004) Constructing visual representations of natural scenes: The roles
of short- and long-term visual memory. J Exp Psychol Hum Percept Perform 30:519 –
537.
Brady et al. www.pnas.org/cgi/content/short/0803390105
1 of 2